search

ZhenglinZhou / STAR

[CVPR 2023] STAR Loss: Reducing Semantic Ambiguity in Facial Landmark Detection
169 stars 20 forks source link

How to convert STAR 98 landmarks to Face alignment 68 landmarks #30

Open Droliven opened 6 months ago

Droliven commented 6 months ago

Thanks for your great job. Could I convert the STAR 98 landmarks to Face alignment 68 landmarks?

ZhenglinZhou commented 6 months ago

Hi @Droliven, thanks for your interest!

Yes, but the corresponding between 98pt (WFLW format) and 68pt (300W format) is not explicit (some discussion can be found here). On the other hand, we also provide the 68pt landmark detector (google / baidu). It can be used in demo.py.

If you have further question, feel free to leave comments.

Droliven commented 6 months ago

Thks, Although we can calculate 68 key points using dlib as shown in the demo.py, I initially chose the STAR model because its network is transparent. I can use STAR to calculate the landmark gradient of my generated image for supervision , but dlib seems to be inconvenient for calculating gradients

ZhenglinZhou commented 6 months ago

I got it.

The absence of a transparent face detector in existing pipeline hinders the further application of STAR? I will consider retraining a new face detector. Thanks for your suggestion!